You are on page 1of 12

R=20190030475 2019-09-14T15:29:28+00:00Z


Quantum Supremacy Using a Programmable

Superconducting Processor

Eleanor G. Rieffel
NASA Ames Research Center

August 2019
NASA STI Program ... in Profile

Since its founding, NASA has been dedicated • CONFERENCE PUBLICATION.

to the advancement of aeronautics and space Collected papers from scientific and
science. The NASA scientific and technical technical conferences, symposia, seminars,
information (STI) program plays a key part in or other meetings sponsored or
helping NASA maintain this important role. co-sponsored by NASA.

The NASA STI program operates under the • SPECIAL PUBLICATION. Scientific,
auspices of the Agency Chief Information Officer. technical, or historical information from
It collects, organizes, provides for archiving, and NASA programs, projects, and missions,
disseminates NASA’s STI. The NASA STI often concerned with subjects having
program provides access to the NTRS Registered substantial public interest.
and its public interface, the NASA Technical
Reports Server, thus providing one of the largest • TECHNICAL TRANSLATION.
collections of aeronautical and space science STI English-language translations of foreign
in the world. Results are published in both non- scientific and technical material pertinent to
NASA channels and by NASA in the NASA STI NASA’s mission.
Report Series, which includes the following report
types: Specialized services also include organizing
and publishing research results, distributing
• TECHNICAL PUBLICATION. Reports of specialized research announcements and
completed research or a major significant feeds, providing information desk and personal
phase of research that present the results of search support, and enabling data exchange
NASA Programs and include extensive data services.
or theoretical analysis. Includes compila-
tions of significant scientific and technical For more information about the NASA STI
data and information deemed to be of program, see the following:
continuing reference value. NASA counter-
part of peer-reviewed formal professional • Access the NASA STI program home page
papers but has less stringent limitations on at
manuscript length and extent of graphic
presentations. • E-mail your question to

• TECHNICAL MEMORANDUM. • Phone the NASA STI Information Desk at

Scientific and technical findings that are 757-864-9658
preliminary or of specialized interest,
e.g., quick release reports, working • Write to:
papers, and bibliographies that contain NASA STI Information Desk
minimal annotation. Does not contain Mail Stop 148
extensive analysis. NASA Langley Research Center
Hampton, VA 23681-2199
• CONTRACTOR REPORT. Scientific and
technical findings by NASA-sponsored
contractors and grantees.

Quantum Supremacy Using a Programmable

Superconducting Processor
Eleanor G. Rieffel
NASA Ames Research Center

National Aeronautics and

Space Administration

Ames Research Center

Moffett Field, California

August 2019

This report is available in electronic form at or
Quantum supremacy using a programmable superconducting processor

Google AI Quantum and collaborators†

The tantalizing promise of quantum computers is that certain computational tasks might be
executed exponentially faster on a quantum processor than on a classical processor. A fundamen-
tal challenge is to build a high-fidelity processor capable of running quantum algorithms in an
exponentially large computational space. Here, we report using a processor with programmable
superconducting qubits to create quantum states on 53 qubits, occupying a state space 253 ∼ 1016 .
Measurements from repeated experiments sample the corresponding probability distribution, which
we verify using classical simulations. While our processor takes about 200 seconds to sample one
instance of the quantum circuit 1 million times, a state-of-the-art supercomputer would require
approximately 10,000 years to perform the equivalent task. This dramatic speedup relative to all
known classical algorithms provides an experimental realization of quantum supremacy on a com-
putational task and heralds the advent of a much-anticipated computing paradigm.

In the early 1980s, Richard Feynman proposed that a DEMONSTRATE QUANTUM SUPREMACY
quantum computer would be an effective tool to solve To demonstrate quantum supremacy, we compare our
problems in physics and chemistry, as it is exponentially quantum processor against state-of-the-art classical com-
costly to simulate large quantum systems with classical puters in the task of sampling the output of a pseudo-
computers [1]. Realizing Feynman’s vision poses signifi- random quantum circuit [24–26]. Random circuits are a
cant experimental and theoretical challenges. First, can suitable choice for benchmarking since they do not pos-
a quantum system be engineered to perform a computa- sess structure and therefore allow for limited guarantees
tion in a large enough computational (Hilbert) space and of computational hardness [24, 25, 27, 28]. We design the
with low enough errors to provide a quantum speedup? circuits to entangle a set of quantum bits (qubits) by re-
Second, can we formulate a problem that is hard for a peated application of single-qubit and two-qubit logical
classical computer but easy for a quantum computer? By operations. Sampling the quantum circuit’s output pro-
computing a novel benchmark task on our superconduct- duces a set of bitstrings, e.g. {0000101, 1011100, . . . }.
ing qubit processor [2–7], we tackle both questions. Our Due to quantum interference, the probability distribution
experiment marks a milestone towards full scale quantum of the bitstrings resembles a speckled intensity pattern
computing: quantum supremacy [8]. produced by light interference in laser scatter, such that
some bitstrings are much more likely to occur than oth-
In reaching this milestone, we show that quantum ers. Classically computing this probability distribution
speedup is achievable in a real-world system and is becomes exponentially more difficult as the number of
not precluded by any hidden physical laws. Quantum qubits (width) and number of gate cycles (depth) grows.
supremacy also heralds the era of Noisy Intermediate-
We verify that the quantum processor is working prop-
Scale Quantum (NISQ) technologies. The benchmark
erly using a method called cross-entropy benchmarking
task we demonstrate has an immediate application in
(XEB) [24, 26], which compares how often each bitstring
generating certifiable random numbers [9]; other initial
is observed experimentally with its corresponding ideal
uses for this new computational capability may include
probability computed via simulation on a classical com-
optimization optimization [10–12], machine learning [13–
puter. For a given circuit, we collect the measured bit-
15], materials science and chemistry [16–18]. However,
strings {xi } and compute the linear XEB fidelity [24–
realizing the full promise of quantum computing (e.g.
26, 29], which is the mean of the simulated probabilities
Shor’s algorithm for factoring) still requires technical
of the bitstrings we measured:
leaps to engineer fault-tolerant logical qubits [19–23].
FXEB = 2n hP (xi )ii − 1 (1)
To achieve quantum supremacy, we made a number of
technical advances which also pave the way towards er- where n is the number of qubits, P (xi ) is the probability
ror correction. We developed fast, high-fidelity gates that of bitstring xi computed for the ideal quantum circuit,
can be executed simultaneously across a two-dimensional and the average is over the observed bitstrings. Intu-
qubit array. We calibrated and benchmarked the pro- itively, FXEB is correlated with how often we sample high
cessor at both the component and system level using a probability bitstrings. When there are no errors in the
powerful new tool: cross-entropy benchmarking (XEB). quantum circuit, sampling the probability distribution
Finally, we used component-level fidelities to accurately will produce FXEB = 1. On the other hand, sampling
predict the performance of the whole system, further from the uniform distribution will give hP (xi )ii = 1/2n
showing that quantum information behaves as expected and produce FXEB = 0. Values of FXEB between 0 and
when scaling to large systems.


We designed a quantum processor named “Sycamore”

which consists of a two-dimensional array of 54 trans-
mon qubits, where each qubit is tunably coupled to four
nearest-neighbors, in a rectangular lattice. The connec-
tivity was chosen to be forward compatible with error-
correction using the surface code [20]. A key systems-
engineering advance of this device is achieving high-
fidelity single- and two-qubit operations, not just in iso-
lation but also while performing a realistic computation
with simultaneous gate operations on many qubits. We
discuss the highlights below; extended details can be
found in the supplementary information.
Qubit Adjustable coupler
In a superconducting circuit, conduction electrons con-
dense into a macroscopic quantum state, such that cur-
b rents and voltages behave quantum mechanically [2, 30].
Our processor uses transmon qubits [6], which can be
thought of as nonlinear superconducting resonators at 5
to 7 GHz. The qubit is encoded as the two lowest quan-
tum eigenstates of the resonant circuit. Each transmon
has two controls: a microwave drive to excite the qubit,
and a magnetic flux control to tune the frequency. Each
qubit is connected to a linear resonator used to read out
the qubit state [5]. As shown in Fig. 1, each qubit is
also connected to its neighboring qubits using a new ad-
justable coupler [31, 32]. Our coupler design allows us to
quickly tune the qubit-qubit coupling from completely
10 millimeters off to 40 MHz. Since one qubit did not function properly
the device uses 53 qubits and 86 couplers.
FIG. 1. The Sycamore processor. a, Layout of processor The processor is fabricated using aluminum for metal-
showing a rectangular array of 54 qubits (gray), each con- ization and Josephson junctions, and indium for bump-
nected to its four nearest neighbors with couplers (blue). In- bonds between two silicon wafers. The chip is wire-
operable qubit is outlined. b, Optical image of the Sycamore bonded to a superconducting circuit board and cooled
chip. to below 20 mK in a dilution refrigerator to reduce am-
bient thermal energy to well below the qubit energy.
The processor is connected through filters and attenu-
1 correspond to the probability that no error has oc- ators to room-temperature electronics, which synthesize
curred while running the circuit. The probabilities P (xi ) the control signals. The state of all qubits can be read
must be obtained from classically simulating the quan- simultaneously by using a frequency-multiplexing tech-
tum circuit, and thus computing FXEB is intractable in nique [33, 34]. We use two stages of cryogenic amplifiers
the regime of quantum supremacy. However, with certain to boost the signal, which is digitized (8 bits at 1 GS/s)
circuit simplifications, we can obtain quantitative fidelity and demultiplexed digitally at room temperature. In to-
estimates of a fully operating processor running wide and tal, we orchestrate 277 digital-to-analog converters (14
deep quantum circuits. bits at 1 GS/s) for complete control of the quantum pro-
Our goal is to achieve a high enough FXEB for a circuit
with sufficient width and depth such that the classical We execute single-qubit gates by driving 25 ns mi-
computing cost is prohibitively large. This is a difficult crowave pulses resonant with the qubit frequency while
task because our logic gates are imperfect and the quan- the qubit-qubit coupling is turned off. The pulses
tum states we intend to create are sensitive to errors. A are shaped to minimize transitions to higher transmon
single bit or phase flip over the course of the algorithm states [35]. Gate performance varies strongly with fre-
will completely shuffle the speckle pattern and result in quency due to two-level-system (TLS) defects [36, 37],
close to 0 fidelity [24, 29]. Therefore, in order to claim stray microwave modes, coupling to control lines and
quantum supremacy we need a quantum processor that the readout resonator, residual stray coupling between
executes the program with sufficiently low error rates. qubits, flux noise, and pulse distortions. We therefore

We model this decay by [1 − e1 /(1 − 1/D2 )]m where e1 is
the Pauli error probability. The state (Hilbert) space di-
mension term, D = 2n = 2, corrects for the depolarizing
Integrated histogram, ECDF

model where states with errors partially overlap with the

ideal state. This procedure is similar to the more typical
technique of randomized benchmarking [21, 38, 39], but
e1 e2 e2c er
supports non-Clifford gatesets [40] and can separate out
decoherence error from coherent control error. We then
repeat the experiment with all qubits executing single-
qubit gates simultaneously (Fig. 2), which shows only a
small increase in the error probabilities, demonstrating
that our device has low microwave crosstalk.
Simultaneous We perform two-qubit iSWAP-like entangling gates by
bringing neighboring qubits on resonance and turning on
Pauli and measurement errors a 20 MHz coupling for 12 ns, which allows the qubits to
Average error Isolated Simultaneous
swap excitations. During this time, the qubits also ex-
Single-qubit (e1) 0.15% 0.16%
perience a controlled-phase (CZ) interaction, which orig-
Two-qubit (e2) 0.36% 0.62% inates from the higher levels of the transmon. The two-
Two-qubit, cycle (e2c) 0.65% 0.93% qubit gate frequency trajectories of each pair of qubits are
Readout (er) 3.1% 3.8% optimized to mitigate the same error mechanisms consid-
ered in optimizing single-qubit operation frequencies.
Pauli error To characterize and benchmark the two-qubit gates,
e1 , e 2 we run two-qubit circuits with m cycles, where each cy-
cle contains a randomly chosen single-qubit gate on each
10-2 of the two qubits followed by a fixed two-qubit gate. We
learn the parameters of the two-qubit unitary (e.g. the
amount of iSWAP and CZ interaction) by using FXEB
as a cost function. After this optimization, we extract
the per-cycle error e2c from the decay of FXEB with m,
and isolate the two-qubit error e2 by subtracting the two
single-qubit errors e1 . We find an average e2 of 0.36%.
Additionally, we repeat the same procedure while simul-
taneously running two-qubit circuits for the entire array.
After updating the unitary parameters to account for ef-
fects such as dispersive shifts and crosstalk, we find an
average e2 of 0.62%.
FIG. 2. System-wide Pauli and measurement errors. a, For the full experiment, we generate quantum circuits
Integrated histogram (empirical cumulative distribution func-
using the two-qubit unitaries measured for each pair dur-
tion, ECDF) of Pauli errors (black, green, blue) and readout
errors (orange), measured on qubits in isolation (dotted lines) ing simultaneous operation, rather than a standard gate
and when operating all qubits simultaneously (solid). The for all pairs. The typical two-qubit gate is a full iSWAP
median of each distribution occurs at 0.50 on the vertical with 1/6 of a full CZ. In principle, our architecture could
axis. Average (mean) values are shown below. b, Heatmap generate unitaries with arbitrary iSWAP and CZ inter-
showing single- and two-qubit Pauli errors e1 (crosses) and e2 actions, but reliably generating a target unitary remains
(bars) positioned in the layout of the processor. Values shown an active area of research.
for all qubits operating simultaneously.
Finally, we benchmark qubit readout using standard
dispersive measurement [41]. Measurement errors aver-
aged over the 0 and 1 states are shown in Fig 2a. We have
optimize the single-qubit operation frequencies to miti- also measured the error when operating all qubits simul-
gate these error mechanisms. taneously, by randomly preparing each qubit in the 0 or
We benchmark single-qubit gate performance by using 1 state and then measuring all qubits for the probability
the XEB protocol described above, reduced to the single- of the correct result. We find that simultaneous readout
qubit level (n = 1), to measure the probability of an error incurs only a modest increase in per-qubit measurement
occurring during a single-qubit gate. On each qubit, we errors.
apply a variable number m of randomly selected gates Having found the error rates of the individual gates and
and measure FXEB averaged over many sequences; as m readout, we can model the fidelity of a quantum circuit
increases, errors accumulate and average FXEB decays. as the product of the probabilities of error-free opera-

a b
single-qubit gate:
25 ns

C XY control

two-qubit gate:
12 ns
B qubit 1
Z control


row A B C D C D A B qubit 2
Z control
cycle: 1 2 3 4 5 6 7 8 m

FIG. 3. Control operations for the quantum supremacy circuits. a, Example quantum circuit instance used in our
√ √ √Every cycle includes a layer each of single- and two-qubit gates. The single-qubit gates are chosen randomly from
{ X, Y , W }. The sequence of two-qubit gates are chosen according to a tiling pattern, coupling each qubit sequentially to
its four nearest-neighbor qubits. The couplers are divided into four subsets (ABCD), each of which is executed simultaneously
across the entire array corresponding to shaded colors. Here we show an intractable sequence (repeat ABCDCDAB); we also
use different coupler subsets along with a simplifiable sequence (repeat EFGHEFGH, not shown) that can be simulated on a
classical computer. b, Waveform of control signals for single- and two-qubit gates.

tion of all gates and measurements. Our largest random feasibility. Finally, we can also run full “verification cir-
quantum circuits have 53 qubits, 1113 single-qubit gates, cuits” with the same gate counts as our supremacy cir-
430 two-qubit gates, and a measurement on each qubit, cuits, but with a different pattern for the sequence of two-
for which we predict a total fidelity of 0.2%. This fidelity qubit gates which is much easier to simulate classically
should be resolvable with a few million
√ measurements, [29]. Comparison between these variations allows track-
since the uncertainty on FXEB is 1/ Ns , where Ns is the ing of the system fidelity as we approach the supremacy
number of samples. Our model assumes that entangling regime.
larger and larger systems does not introduce additional We first check that the patch and elided versions of the
error sources beyond the errors we measure at the single- verification circuits produce the same fidelity as the full
and two-qubit level — in the next section we will see how verification circuits up to 53 qubits, as shown in Fig. 4a.
well this hypothesis holds. For each data point, we typically collect Ns = 5 × 106
total samples over ten circuit instances, where instances
FIDELITY ESTIMATION IN THE SUPREMACY differ only in the choices of single-qubit gates in each
REGIME cycle. We also show predicted FXEB values computed
The gate sequence for our pseudo-random quantum by multiplying the no-error probabilities of single- and
circuit generation is shown in Fig. 3. One cycle of the two-qubit gates and measurement [29]. Patch, elided,
algorithm consists√of applying
√ √ single-qubit gates chosen and predicted fidelities all show good agreement with
randomly from { X, Y , W } on all qubits, followed the fidelities of the corresponding full circuits, despite
by two-qubit gates on pairs of qubits. The sequences of the vast differences in computational complexity and en-
gates which form the “supremacy circuits” are designed tanglement. This gives us confidence that elided circuits
to minimize the circuit depth required to create a highly can be used to accurately estimate the fidelity of more
entangled state, which ensures computational complexity complex circuits.
and classical hardness. We proceed now to benchmark our most computa-
While we cannot compute FXEB in the supremacy tionally difficult circuits. In Fig. 4b, we show the mea-
regime, we can estimate it using three variations to re- sured FXEB for 53-qubit patch and elided versions of the
duce the complexity of the circuits. In “patch circuits”, full supremacy circuits with increasing depth. For the
we remove a slice of two-qubit gates (a small fraction largest circuit with 53 qubits and 20 cycles, we collected
of the total number of two-qubit gates), splitting the cir- Ns = 30×106 samples over 10 circuit instances, obtaining
cuit into two spatially isolated, non-interacting patches of FXEB = (2.24 ± 0.21) × 10−3 for the elided circuits. With
qubits. We then compute the total fidelity as the product 5σ confidence, we assert that the average fidelity of run-
of the patch fidelities, each of which can be easily calcu- ning these circuits on the quantum processor is greater
lated. In “elided circuits”, we remove only a fraction of than at least 0.1%. The full data for Fig. 4b should have
the initial two-qubit gates along the slice, allowing for similar fidelities, but are only archived since the simula-
entanglement between patches, which more closely mim- tion times (red numbers) take too long. It is thus in the
ics the full experiment while still maintaining simulation quantum supremacy regime.

a Classically verifiable b Supremacy regime



Sycamore sampling (Ns = 1M): 200 seconds

2 hours Classical sampling @ 𝓕Sycamore

2 weeks
XEB Fidelity,

Classical verification 1 week

4 years
4 years
5 hours 100 years
600 years
10 millennia
10-2 m = 14 cycles
Prediction from gate and measurement errors
Full circuit Elided circuit Patch circuit

n = 53 qubits
Elided (±5σ error bars)
10 15 20 25 30 35 40 45 50 55 12 14 16 18 20
number of qubits, n number of cycles, m

FIG. 4. Demonstrating quantum supremacy. a, Verification of benchmarking methods. FXEB values for patch, elided,
and full verification circuits are calculated from measured bitstrings and the corresponding probabilities predicted by classical
simulation. Here, the two-qubit gates are applied in a simplifiable tiling and sequence such that the full circuits can be simulated
out to n = 53, m = 14 in a reasonable amount of time. Each data point is an average over 10 distinct quantum circuit instances
that differ in their single-qubit gates (for n = 39, 42, 43 only 2 instances were simulated). For each n, each instance is sampled
with Ns between 0.5 M and 2.5 M. The black line shows predicted FXEB based on single- and two-qubit gate and measurement
errors. The close correspondence between all four curves, despite their vast differences in complexity, justifies the use of elided
circuits to estimate fidelity in the supremacy regime. b, Estimating FXEB in the quantum supremacy regime. Here, the
two-qubit gates are applied in a non-simplifiable tiling and sequence for which it is much harder to simulate. For the largest
elided data (n = 53, m = 20, total Ns = 30 M), we find an average FXEB > 0.1% with 5σ confidence, where σ includes both
systematic and statistical uncertainties. The corresponding full circuit data, not simulated but archived, is expected to show
similarly significant fidelity. For m = 20, obtaining 1M samples on the quantum processor takes 200 seconds, while an equal
fidelity classical sampling would take 10,000 years on 1M cores, and verifying the fidelity would take millions of years.

DETERMINING THE CLASSICAL the Feynman path-integral. While it is more memory-

COMPUTATIONAL COST efficient, SFA becomes exponentially more computation-
ally expensive with increasing circuit depth due to the
We simulate the quantum circuits used in the exper- exponential growth of paths with the number of gates
iment on classical computers for two purposes: verify- connecting the patches.
ing our quantum processor and benchmarking methods
by computing FXEB where possible using simplifiable To estimate the classical computational cost of the
circuits (Fig. 4a), and estimating FXEB as well as the supremacy circuits (gray numbers, Fig. 4b), we ran por-
classical cost of sampling our hardest circuits (Fig. 4b). tions of the quantum circuit simulation on both the Sum-
Up to 43 qubits, we use a Schrödinger algorithm (SA) mit supercomputer as well as on Google clusters and ex-
which simulates the evolution of the full quantum state; trapolated to the full cost. In this extrapolation, we
the Jülich supercomputer (100k cores, 250TB) runs the account for the computational cost scaling with FXEB ,
largest cases. Above this size, there is not enough RAM e.g. the 0.1% fidelity decreases the cost by 1000 [43, 44].
to store the quantum state [42]. For larger qubit num- On the Summit supercomputer, which is currently the
bers, we use a hybrid Schrödinger-Feynman algorithm most powerful in the world, we used a method inspired
(SFA) [43] running on Google data centers to compute by Feynman path-integrals that is most efficient at low
the amplitudes of individual bitstrings. This algorithm depth [44–47]. At m = 20 the tensors do not reasonably
breaks the circuit up into two patches of qubits and effi- fit in node memory, so we can only measure runtimes
ciently simulates each patch using a Schrödinger method, up to m = 14, for which we estimate that sampling 3M
before connecting them using an approach reminiscent of bitstrings with 1% fidelity would require 1 year.

On Google Cloud servers, we estimate that perform- tum processors have thus reached the regime of quantum
ing the same task for m = 20 with 0.1% fidelity using supremacy. We expect their computational power will
the SFA algorithm would cost 50 trillion core-hours and continue to grow at a double exponential rate: the clas-
consume 1 petawatt hour of energy. To put this in per- sical cost of simulating a quantum circuit increases expo-
spective, it took 600 seconds to sample the circuit on nentially with computational volume, and hardware im-
the quantum processor 3 million times, where sampling provements will likely follow a quantum-processor equiv-
time is limited by control hardware communications; in alent of Moore’s law [52, 53], doubling this computational
fact, the net quantum processor time is only about 30 volume every few years. To sustain the double exponen-
seconds. The bitstring samples from this largest circuit tial growth rate and to eventually offer the computational
are archived online. volume needed to run well-known quantum algorithms,
One may wonder to what extent algorithmic innova- such as the Shor or Grover algorithms [19, 54], the engi-
tion can enhance classical simulations. Our assumption, neering of quantum error correction will have to become
based on insights from complexity theory, is that the cost a focus of attention.
of this algorithmic task is exponential in n as well as m. The “Extended Church-Turing Thesis” formulated by
Indeed, simulation methods have improved steadily over Bernstein and Vazirani [55] asserts that any “reasonable”
the past few years [42–50]. We expect that lower simula- model of computation can be efficiently simulated by a
tion costs than reported here will eventually be achieved, Turing machine. Our experiment suggests that a model
but we also expect they will be consistently outpaced by of computation may now be available that violates this
hardware improvements on larger quantum processors. assertion. We have performed random quantum circuit
sampling in polynomial time with a physically realized
quantum processor (with sufficiently low error rates), yet
no efficient method is known to exist for classical comput-
A key assumption underlying the theory of quantum ing machinery. As a result of these developments, quan-
error correction is that quantum state errors may be con- tum computing is transitioning from a research topic to a
sidered digitized and localized [38, 51]. Under such a dig- technology that unlocks new computational capabilities.
ital model, all errors in the evolving quantum state may We are only one creative algorithm away from valuable
be characterized by a set of localized Pauli errors (bit- near-term applications.
and/or phase-flips) interspersed into the circuit. Since
continuous amplitudes are fundamental to quantum me-
chanics, it needs to be tested whether errors in a quantum
system could be treated as discrete and probabilistic. In-
deed, our experimental observations support the validity Acknowledgments We are grateful to Eric Schmidt,
Sergey Brin, Jeff Dean, and Jay Yagnik for their executive
of this model for our processor. Our system fidelity is
sponsorship of the Google AI Quantum team, and for their
well predicted by a simple model in which the individ- continued engagement and support. We thank Peter Norvig
ually characterized fidelities of each gate are multiplied for reviewing a draft of the manuscript, and Sergey Knysh
together (Fig 4). for useful discussions. We thank Kevin Kissel, Joey Raso,
To be successfully described by a digitized error model, Davinci Yonge-Mallo, Orion Martin, and Niranjan Sridhar
a system should be low in correlated errors. We achieve for their help with simulations. We thank Gina Bortoli and
this in our experiment by choosing circuits that ran- Lily Laws for keeping our team organized. This research used
resources from the Oak Ridge Leadership Computing Facility,
domize and decorrelate errors, by optimizing control to
which is a DOE Office of Science User Facility supported un-
minimize systematic errors and leakage, and by design- der Contract DE-AC05-00OR22725. A portion of this work
ing gates that operate much faster than correlated noise was performed in the UCSB Nanofabrication Facility, an open
sources, such as 1/f flux noise [37]. Demonstrating a pre- access laboratory.
dictive uncorrelated error model up to a Hilbert space of
Author contributions The Google AI Quantum team
size 253 shows that we can build a system where quantum conceived of the experiment. The applications and algorithms
resources, such as entanglement, are not prohibitively team provided the theoretical foundation and the specifics of
fragile. the algorithm. The hardware team carried out the experiment
and collected the data. The data analysis was done jointly
with outside collaborators. All authors wrote and revised the
WHAT DOES THE FUTURE HOLD? manuscript and supplement.
Quantum processors based on superconducting qubits Competing Interests The authors declare that they have
can now perform computations in a Hilbert space of di- no competing financial interests.
mension 253 ≈ 9 × 1015 , beyond the reach of the fastest Correspondence and requests for materials should
classical supercomputers available today. To our knowl- be addressed to John M. Martinis (
edge, this experiment marks the first computation that
can only be performed on a quantum processor. Quan-

Frank Arute1 , Kunal Arya1 , Ryan Babbush1 , Dave on Physics, Brussels (2012).
Bacon1 , Joseph C. Bardin1,2 , Rami Barends1 , Ru- [9] Aaronson, S. Certified randomness from quantum
pak Biswas3 , Sergio Boixo1 , Fernando G.S.L. Brandao1,4 , supremacy. In preparation .
David Buell1 , Brian Burkett1 , Yu Chen1 , Zijun Chen1 , [10] Hastings, M. B. Classical and Quantum Bounded
Ben Chiaro5 , Roberto Collins1 , William Courtney1 , An- Depth Approximation Algorithms. arXiv e-prints
drew Dunsworth1 , Edward Farhi1 , Brooks Foxen5 , Austin arXiv:1905.07047 (2019). 1905.07047.
Fowler1 , Craig Gidney1 , Marissa Giustina1 , Rob Graff1 , Keith [11] Kechedzhi, K. et al. Efficient population transfer via non-
Guerin1 , Steve Habegger1 , Matthew P. Harrigan1 , Michael J. ergodic extended states in quantum spin glass. arXiv
Hartmann1,6 , Alan Ho1 , Markus Hoffmann1 , Trent Huang1 , e-prints arXiv:1807.04792 (2018). 1807.04792.
Travis S. Humble7 , Sergei V. Isakov1 , Evan Jeffrey1 , Zhang [12] Somma, R. D., Boixo, S., Barnum, H. & Knill, E. Quan-
Jiang1 , Dvir Kafri1 , Kostyantyn Kechedzhi1 , Julian Kelly1 , tum simulations of classical annealing processes. Phys.
Paul V. Klimov1 , Alexander Korotkov1 , Fedor Kostritsa1 , Rev. Lett. letters 101, 130504 (2008).
David Landhuis1 , Mike Lindmark1 , Erik Lucero1 , Dmitry [13] McClean, J. R., Boixo, S., Smelyanskiy, V. N., Babbush,
Lyakh7 , Salvatore Mandrá3 , Jarrod R. McClean1 , Matt R. & Neven, H. Barren plateaus in quantum neural net-
McEwen5 ,Anthony Megrant1 , Xiao Mi1 ,Kristel Michielsen8 , work training landscapes. Nat. Comm. 9, 4812 (2018).
Masoud Mohseni1 , Josh Mutus1 , Ofer Naaman1 , Matthew [14] Cong, I., Choi, S. & Lukin, M. D. Quantum convolutional
Neeley1 , Charles Neill1 , Murphy Yuezhen Niu1 , Eric Ostby1 , neural networks. arXiv:1810.03787 (2018).
Andre Petukhov1 , John C. Platt1 , Chris Quintana1 , Eleanor [15] Bravyi, S., Gosset, D. & König, R. Quantum advantage
G. Rieffel3 , Pedram Roushan1 , Nicholas Rubin1 , Daniel with shallow circuits. Science 362, 308–311 (2018).
Sank1 , Kevin J. Satzinger1 , Vadim Smelyanskiy1 , Kevin [16] Aspuru-Guzik, A., Dutoi, A. D., Love, P. J. & Head-
Sung1 , Matthew D. Trevithick1 , Amit Vainsencher1 , Ben- Gordon, M. Simulated quantum computation of molecu-
jamin Villalonga1,9 , Theodore White1 , Jamie Yao1 , Ping lar energies. Science 309, 1704–1707 (2005).
Yeh1 , Adam Zalcman1 , Hartmut Neven1 , John M. Martinis1,5 [17] Peruzzo, A. et al. A variational eigenvalue solver on a
photonic quantum processor. Nat. Commun. 5, 4213
1. Google Research, Mountain View, CA 94043, USA, 2. (2014).
Department of Electrical and Computer Engineering, Uni- [18] Hempel, C. et al. Quantum chemistry calculations on a
versity of Massachusetts Amherst, Amherst, MA, USA, 3. trapped-ion quantum simulator. Phys. Rev. X 8, 031022
Quantum Artificial Intelligence Lab. (QuAIL), NASA Ames (2018).
Research Center, Moffett Field, USA, 4. Institute for [19] Shor, P. W. Algorithms for quantum computation: dis-
Quantum Information and Matter, Caltech, Pasadena, CA, crete logarithms and factoring proceedings. Proceedings
USA, 5. Department of Physics, University of California, 35th Annual Symposium on Foundations of Computer
Santa Barbara, CA, USA, 6. Friedrich-Alexander University Science (1994).
Erlangen-Nürnberg (FAU), Department of Physics, Erlangen, [20] Fowler, A. G., Mariantoni, M., Martinis, J. M. & Cle-
Germany, 7. Quantum Computing Institute, Oak Ridge Na- land, A. N. Surface codes: Towards practical large-scale
tional Laboratory, Oak Ridge, TN, USA, 8. Institute for quantum computation. Phys. Rev. A 86, 032324 (2012).
Advanced Simulation, Jülich Supercomputing Centre, Jülich, [21] Barends, R. et al. Superconducting quantum circuits at
Germany, 9. Department of Physics, University of Illinois the surface code threshold for fault tolerance. Nature
at Urbana-Champaign, Urbana, IL, USA 508, 500–503 (2014).
[22] Córcoles, A. D. et al. Demonstration of a quantum error
detection code using a square lattice of four supercon-
ducting qubits. Nat. Commun. 6, 6979 (2015).
[23] Ofek, N. et al. Extending the lifetime of a quantum bit
[1] Feynman, R. P. Simulating physics with computers. Int. with error correction in superconducting circuits. Nature
J. Theor. Phys. 21, 467–488 (1982). 536, 441 (2016).
[2] Devoret, M. H., Martinis, J. M. & Clarke, J. Mea- [24] Boixo, S. et al. Characterizing quantum supremacy in
surements of macroscopic quantum tunneling out of the near-term devices. Nat. Phys. 14, 595 (2018).
zero-voltage state of a current-biased josephson junction. [25] Aaronson, S. & Chen, L. Complexity-theoretic founda-
Phys. Rev. Lett 55, 1908 (1985). tions of quantum supremacy experiments. In 32nd Com-
[3] Nakamura, Y., Chen, C. D. & Tsai, J. S. Spectroscopy putational Complexity Conference (CCC 2017) (2017).
of energy-level splitting between two macroscopic quan- [26] Neill, C. et al. A blueprint for demonstrating quantum
tum states of charge coherently superposed by josephson supremacy with superconducting qubits. Science 360,
coupling. Phys. Rev. Lett. 79, 2328 (1997). 195–199 (2018).
[4] Mooij, J. et al. Josephson persistent-current qubit. Sci- [27] Bremner, M. J., Montanaro, A. & Shepherd, D. J.
ence 285, 1036 (1999). Average-case complexity versus approximate simulation
[5] Wallraff, A. et al. Strong coupling of a single photon to a of commuting quantum computations. Phys. Rev. Lett.
superconducting qubit using circuit quantum electrody- 117, 080501 (2016).
namics. Nature 431, 162 (2004). [28] Bouland, A., Fefferman, B., Nirkhe, C. & Vazi-
[6] Koch, J. et al. Charge-insensitive qubit design derived rani, U. Quantum supremacy and the com-
from the cooper pair box. Phys. Rev. A 76, 042319 plexity of random circuit sampling. Preprint at
(2007). (2018).
[7] You, J. Q. & Nori, F. Atomic physics and quantum optics [29] See supplementary information .
using superconducting circuits. Nature 474, 589 (2011). [30] Vool, U. & Devoret, M. Introduction to quantum electro-
[8] Preskill, J. Quantum computing and the entanglement magnetic circuits. Int. J. Circ. Theor. Appl. 45, 897–934
frontier. Rapporteur talk at the 25th Solvay Conference (2017).

[31] Chen, Y. et al. Qubit architecture with high coherence [44] Villalonga, B. et al. A flexible high-performance sim-
and fast tunable coupling circuits. Phys. Rev. Lett. 113, ulator for the verification and benchmarking of quan-
220502 (2014). tum circuits implemented on real hardware. Preprint at
[32] Yan, F. et al. A tunable coupling scheme for implement- (2018).
ing high-fidelity two-qubit gates. Phys. Rev. Applied 10, [45] Boixo, S., Isakov, S. V., Smelyanskiy, V. N. &
054062 (2018). Neven, H. Simulation of low-depth quantum circuits
[33] Schuster, D. I. et al. Resolving photon number states in as complex undirected graphical models. Preprint at
a superconducting circuit. Nature 445, 515 (2007). (2017).
[34] Jeffrey, E. et al. Fast accurate state measurement with [46] Chen, J., Zhang, F., Huang, C., Newman, M. & Shi,
superconducting qubits. Phys. Rev. Lett. 112, 190504 Y. Classical simulation of intermediate-size quantum
(2014). circuits. Preprint at
[35] Chen, Z. et al. Measuring and suppressing quantum state (2018).
leakage in a superconducting qubit. Phys. Rev. Lett. 116, [47] Villalonga, B. et al. Establishing the quantum supremacy
020501 (2016). frontier with a 281 pflop/s simulation. Preprint at
[36] Klimov, P. V. et al. Fluctuations of energy-relaxation (2019).
times in superconducting qubits. Phys. Rev. Lett. 121, [48] Pednault, E. et al. Breaking the 49-qubit barrier
090502 (2018). in the simulation of quantum circuits. Preprint at
[37] Yan, F. et al. The flux qubit revisited to enhance coher- (2017).
ence and reproducibility. Nat. Commun. 7, 12964 (2016). [49] Chen, Z. Y. et al. 64-qubit quantum circuit simulation.
[38] Knill, E. et al. Randomized benchmarking of quantum Sci. Bull. 63, 964–971 (2018).
gates. Phys. Rev. A 77, 012307 (2008). [50] Chen, M.-C. et al. Quantum teleportation-inspired al-
[39] Magesan, E., Gambetta, J. M. & Emerson, J. Scalable gorithm for sampling large random quantum circuits.
and robust randomized benchmarking of quantum pro- Preprint at (2019).
cesses. Phys. Rev. Lett. 106, 180504 (2011). [51] Shor, P. W. Scheme for reducing decoherence in quan-
[40] Cross, A. W., Magesan, E., Bishop, L. S., Smolin, J. A. & tum computer memory. Phys. Rev. A 52, R2493–R2496
Gambetta, J. M. Scalable randomised benchmarking of (1995).
non-clifford gates. NPJ Quantum Information 2, 16012 [52] Devoret, M. H. & Schoelkopf, R. J. Superconducting
(2016). circuits for quantum information: An outlook. Science
[41] Wallraff, A. et al. Approaching unit visibility for control 339, 1169–1174 (2013).
of a superconducting qubit with dispersive readout. Phys. [53] Mohseni, M. et al. Commercialize quantum technologies
Rev. Lett. 95, 060501 (2005). in five years. Nature 543, 171 (2017).
[42] De Raedt, H. et al. Massively parallel quantum computer [54] Grover, L. K. Quantum mechanics helps in searching for
simulator, eleven years later. Comput. Phys. Commun. a needle in a haystack. letters 79, 325 (1997).
237, 47 – 61 (2019). [55] Bernstein, E. & Vazirani, U. Quantum complexity the-
[43] Markov, I. L., Fatima, A., Isakov, S. V. & Boixo, S. ory. Proc. 25th Annual ACM Symposium on Theory of
Quantum supremacy is both closer and farther than it Computing (1993).
appears. Preprint at